YT YouTube2Text

YouTube2Text → Transcripts → My top secrets to running an AI Agent Workforce

My top secrets to running an AI Agent Workforce — Transcript

by Greg Isenberg · 8,835 words · 1,351 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00There are people that are spitting up
0:01agent workforces with hundreds of agents
0:04and sub agents and they're getting
0:06incredible amounts of work [music] done.
0:08But how do you do it? And how could you
0:11think about it? And what are the
0:12strategies to actually create an AI
0:14agent workforce
0:15>> [music]
0:16>> that under promises and over delivers?
0:18Well, today I brought on Ali K. Miller
0:21and she's one of one of the most
0:22well-known AI voices ever. She's worked
0:25with IBM, she's worked with AWS and
0:28she's managed multi-billion dollar P&Ls
0:30in the AI space.
0:32I asked her a simple question, how do
0:34you manage your fleet of agents? In this
0:37episode, we cover a lot of ground, but
0:39by the end of it, you're going to
0:40understand how should you should
0:42strategically think about spinning up AI
0:45agent workforces, where there's
0:47opportunities to create startups in the
0:49B2B space with AI agents and a lot
0:52[music] more. Enjoy the episode and I'll
0:54see you at the end. Today's episode is
0:56brought to you by Brex. My company's
0:58been on Brex for a year and a half and I
1:00started because I kept hearing companies
1:02like Vercel, OpenAI, and Anthropic were
1:04using Brex and I figured if they're
1:06using it, why shouldn't I? It's been a
1:09game-changer. The thing that got me is
1:11how smooth it is. It's got high-limit
1:13cards, it's got banking, it's got AI
1:16that handles the back office busywork
1:18like expense reports, which I don't want
1:19to do, on its own. It's really just
1:21built for this agentic world. If you're
1:24building something new, it's time to get
1:26Brex. Check it out at
1:27brex.com/solutions/startups.
1:30Link in the description.
1:35>> [music]
1:40>> I can't tell you how excited I am to
1:42finally have Ali Miller on the podcast.
1:46I've been begging her to come on. She's
1:48one of my favorite people in AI and I
1:50don't say that lightly. Um welcome to
1:53the show, Ali.
1:55>> Thank you, Greg. And you are also one of
1:56my favorite people, so like I'm actually
1:59very excited to to talk about all the AI
2:01things that we're working on.
2:03>> By the end of the episode, what are
2:04people going to learn?
2:05>> I hope like one of the biggest mindset
2:08shifts that I'm going through right now
2:09is I feel like the term managing agents
2:12is wrong. And my hope is that people
2:16will understand what that mindset shift
2:18is, see a few examples, and figure out
2:21how to start how to make that mindset
2:24shift, what the first step should be.
2:26>> Okay, perfect. So, where do you want to
2:28start?
2:29>> So, this is and and I'm happy to to
2:31debate you on this cuz we haven't
2:33chatted about this. But I feel like
2:37managing agents feels like I'm their
2:39direct manager and I'm like, "Suzy, go
2:41over there and Betty, go over there and
2:43Jeremy, go over here." And I feel like I
2:46am three rungs above at like an SVP
2:50overseeing level
2:52where I feel like I am setting up the
2:55infrastructure and then
2:58they are figuring out the best way to
3:00execute within that. Um and so I feel
3:03like I'm moving from managing to like
3:05waiting for escalations.
3:07Um or I feel like I'm moving away from
3:09delegating and more just deciding what
3:13should or shouldn't happen. And so it's
3:15a little bit more of the like a like a
3:17liability role where I just get to be
3:20the the final say of what happens
3:23um and come in for like critical
3:24thinking stops. But does it like am I
3:27the only one that feels like that is
3:29happening? I just it feels like that
3:31word is wrong. Like I see managing
3:33agents everywhere and it just feels like
3:35anyone that is still talking about, "You
3:37should manage agents." feels like early
3:392026 talk.
3:40>> Also, like do do we want to manage
3:43agents? Is also the question. Like
3:45managing people is hard, you know what I
3:47mean? Like
3:49a big reason I think a lot of people
3:51like AI to do stuff for us is so we
3:55don't have to manage things, you know?
3:58So, that's something else I've been
3:59thinking about.
4:00>> Like, I I ran an org of about 100 people
4:03at AWS. The parts of people management
4:06that I loved, it was
4:09the the making them better and
4:12empowering the out of them and
4:15seeing them completely blow past their
4:18ceiling, watching them get promotions.
4:20Like, that was the fun part and also
4:22seeing what we could do together. Things
4:24like, "Oh, we have to fill out this
4:28thing with the paper and the button."
4:29And like, get me out of there. So, I
4:31think the admin side of people
4:33management and the admin side of agent
4:35management, I want that fully gone.
4:38The things that I that I love about
4:40people, I'm bringing that over into
4:41agents, which is just like, "How do I
4:44act as as ambitiously as possible and
4:47get you to break through your ceiling?"
4:49And one of the best prompts that I have
4:51done with my AI workforce is three
4:55words.
4:56>> [laughter]
4:57>> And with like a little bit of
4:59explanation, but like, at its core, it
5:01is three words. That is the best prompt
5:02ever. So, I have
5:05uh my AI chief of staff is Simon. Simon
5:07runs like this whole org. And so, I have
5:0934 AI agents that work in this
5:10workforce.
5:11And it dawned on me
5:13that I was already functioning at the
5:17limit of my own imagination in my
5:18business. And that I could be doing way
5:20more ambitious things if only someone
5:23could manage me, right? Like, could
5:25break help me break through my ceiling.
5:27And obviously, I have a lot of mentors
5:28and you're amazing at at, you know,
5:30shaking people up and and making me
5:32second guess how I'm doing things. It's
5:33really helpful. But, I it dawned on me.
5:36I was like, "Why am I not leaning on the
5:39AI agents to help me with this? Like,
5:41why is everything that they're working
5:43on initially prompted by me? Even if
5:46it's
5:47um a scheduled task, I still had to come
5:49up with that task and tell it to do it.
5:52So, the best prompt, three words, and
5:54it's just do smart things.
5:57Like, my AI workforce has access to
6:00every single context doc I've got.
6:02Context docs about my business, my
6:03friends, family, my 2026 personal goals,
6:07business goals. It has access to my
6:09meeting transcripts, email, calendar,
6:11Notion, Stripe, Supabase, GitHub,
6:14whatever.
6:15And I just several times a day want it
6:19to look across all these things and just
6:21do smart things.
6:23And seeing how Fable 5 and GPT 5.6 and
6:27that level model is reacting to that
6:30vague
6:32um flavor of prompt. Like, you could you
6:35couldn't do this a year ago. Now, you
6:37absolutely can.
6:38>> So, when you hire a human being, I think
6:41there's like three types of employees
6:44that you can have. One is uh someone who
6:47doesn't complete tasks, not a good
6:49employee if they're not completing
6:50tasks. Um the second is they're
6:53completing the tasks um like
6:56satisfactory or exceeding, but like
6:58they're not really like thinking about
6:59new tasks. Um so, they're not You can't
7:03just like if you step away from the
7:05business, you're probably not going to
7:07see insane growth. Um
7:09and then the best employee that you can
7:11possibly hire is doing the task,
7:13exceeding expectations on it, but also
7:16thinking about new tasks that they
7:18should be doing, and actually going and
7:20doing those, and
7:21exceeding expectations or or you know,
7:24or being very satisfactory on that. So,
7:25what you're saying is
7:27basically, you're just giving more
7:29responsibility to your team of agents.
7:32Um you're giving in a way cuz you're
7:33giving these three words to it and
7:35you're saying like, "Hey, I'm shifting
7:36the responsibility of like
7:39you know, do smart things to you." Like,
7:42you have to you you have to like there's
7:44a bunch of fog that you have to figure
7:45out.
7:46>> Yes. I would say I'm giving them more
7:47breath, more scope, more flexibility.
7:50I'm not allowing them to now send 100
7:54emails and before I used to have to
7:56check all the emails. I still check all
7:57the emails. So, the the tier of risk has
8:00stayed the same, but the width has
8:03expanded.
8:04>> It's it's almost like unbelievable that
8:07those three words actually make a
8:08difference.
8:09>> Yes. This is like I And by the way, so
8:12so I I agree with your assessment on
8:15this like tiers of employees and Alex
8:19Lieberman shared this like pyramid of
8:20proactivity that I turned into I'll I'll
8:23send this to you so that you can pull it
8:24up right now as I'm talking about it.
8:25But, it is five levels of proactivity.
8:28And at level four, it's like I've
8:32already solved this thing. Here are the,
8:35you know, tradeoffs or whatever. And at
8:38level five, it's like I've already
8:40solved this thing. Here's how I'm going
8:42to deal with it if it goes wrong. Here's
8:44the next steps, all the things that you
8:45just laid out. I would say that the
8:47difference between someone who's at
8:48level three and two in in your um
8:51analogy is someone that understands
8:53goals and someone who's been given the
8:56power
8:57to rethink how things get done and the
9:00power to actually execute. And I give my
9:04AI workforce goals. Like, those are
9:06written out and every single quarter
9:09also um share with you this tweet that
9:11has like the prompt that I think
9:12everyone can use. But, every single
9:14quarter I'm going through a goals review
9:16with my AI agent workforce so that the
9:19goals documents that are living on my
9:21desktop and are duplicated in the drive
9:23so that all this cloud like workflows
9:25can actually work. Um all of that is so
9:30that AI, when it is in that expanded
9:33scope world and it's taking on that new
9:35tasks,
9:36it's doing it in a goal-oriented way.
9:38It's like giving it a product mindset.
9:41Like I think it would be extremely
9:43limiting if you only treated this thing
9:45as an engineer when it could be the
9:47greatest product lead you've ever had.
9:50>> I think you tweeted about like your your
9:52like men you're really focused on
9:54proactive agents, right?
9:56>> Yes.
9:57>> about when when you talk about proactive
9:58agents? Is this what you're talking
10:00about?
10:01>> So I when you talk to the AI labs and I
10:03know you do and I know I do and a bunch
10:05of others probably do. But the the word
10:07of the year feels like it's proactive.
10:10So I don't want to be the first domino
10:13anymore. I don't want to be the
10:15bottleneck in my own work and any single
10:18moment that I realize that I am the
10:20limiting factor of helping a billion
10:23people transform their lives, work, and
10:26business in the AI age, I have to remove
10:28myself from the process and go, "Bad
10:30alley, like what are you doing?"
10:32[laughter]
10:33And so a lot of that um especially in
10:35the in the kind of tail end of 2025,
10:38first half of 2026 was switching into
10:42proactive agents. So we we had proactive
10:44automations that were trigger-based. Um
10:47I'll give you a really easy example.
10:50Every single time I drop a video
10:53recording into our video folder, so
10:54basically anytime I do a screen
10:56recording, goes into this one folder and
10:58automatically it gets generated um
11:01automatically generated is a transcript
11:02of that video um that gets, you know,
11:05then saved into our little transcripty
11:07thing. Social posts get generated that
11:10are in my voice, so nine different
11:11social posts get generated for X and
11:14LinkedIn and Instagram real scripts and
11:16all this stuff. So that presumably the
11:18thing that I was filming was for a
11:20social video. Um so that was easy
11:24automation land, but that is just one
11:26example of like a proactive um
11:29very
11:30um
11:32well-defined workflow.
11:34What I think is more interesting for the
11:37back half of 2026 is proactive of
11:40undefined workflows. So, like AI is
11:42probabilistic all the time and not
11:45deterministic, but I want to take that
11:48probabilistic nature of reasoning, like
11:50the step zero of reasoning, and apply
11:52that to the actual tasks that it takes
11:54on. So, in order to do that, whether
11:56you're talking to a human or an agent,
11:58they have to know what's the goal,
11:59what's the star, what's that vision.
12:02They have to have access to tools,
12:03permission to use these tools in the way
12:05that actually gets work off your plate,
12:07and a sense of what would normally
12:10trigger that sort of action.
12:12So, and I can I'm going to share
12:15one thing [clears throat]
12:16on on screen here, which is every single
12:20day,
12:21um let me just give me 1 second.
12:24So, essentially, like
12:27I want my whole company to be queryable.
12:30I want AI to have context on everything
12:33that's happening, and it dawned on me
12:35that yes, it had access to all my
12:36meeting transcripts, and it had access
12:39to my Gmail and all this stuff, but
12:41there was a lot that was not yet
12:43codified, and it was things like
12:46everything is becoming proactive, I want
12:48to be more proactive, or
12:51this client, they think that what they
12:53need help with is workflows, you know,
12:55under the CMO, but actually what they
12:57have problems with is reskilling and
13:00finding new roles for this one
13:01department. So, anything that is not
13:03codified inside of, again, meetings,
13:05emails, whatever, or Slack, I have asked
13:09AI now to prompt me every single day
13:11with this, and you know, I got to put it
13:13in my brand colors, and I didn't want to
13:16have to think with, you know, maybe 10%
13:19of my brain still working at the end of
13:20the day, so I give it like a little
13:22prompt. It reminds me to dictate because
13:24that's four times faster than writing.
13:26And so, I will bank these entries to be
13:29like
13:30you know, I talked to Greg. I feel like
13:31the entire focus is on proactive agents,
13:33proactivity,
13:35um and flexibility. And I want to look
13:38more into his three levels of employees.
13:41And so, like I might do this for 5
13:44minutes or 40 minutes at the end of at
13:46the end of the day. I might do it
13:47throughout the day. And then I just save
13:48it out and then it's like it this goes
13:51into my personal wiki. And all I want to
13:56do
13:57is make sure that the agents that are
13:59working at that really flexible layer
14:02where again, I am not managing them.
14:05I am enabling them and they're coming
14:07back to me with those escalations and
14:09decisions.
14:11I want to make sure that they have the
14:12right context or else all their stuff is
14:14going to be wrong. And and we saw this
14:16in the beginning of our AI workforce
14:17stuff. It was like, oh, I saw that, you
14:20know, Greg confirmed that interview. And
14:22it's like, no, Greg confirmed it, but
14:24we're still figuring out dates and I'm
14:25doing it over text and you know,
14:27IMessage MCP broke since you can't see
14:29that. So, there was a lot of stuff that
14:31we had to continually fix and it took
14:34probably months to get to where we are
14:35now. But, we have Claude in every single
14:38one of our chat channels. I had a very
14:40weird I I have to send I have to show
14:42you this.
14:43Um
14:45Let me just share my whole screen.
14:46>> By the way, this So, the Brain Meets
14:49Diary thing, so when you
14:51>> Yeah.
14:51>> when you you know, you add today Well,
14:55you had like 86 entries, right? So, your
14:58AI agents do all of your Does your Does
15:01your entire AI workforce workforce have
15:04access to that or just some? How do you
15:06think about that?
15:08>> So, great question.
15:11Um essentially, my AI workforce right
15:13now is one AI chief of staff with six
15:16directors. Those directors are largely
15:19over like business functions. So, one is
15:21education, one is all the client work,
15:24um one is kind of operations, one's
15:25marketing, one product, and then Phoebe,
15:28all these are named after Friends
15:29characters. Phoebe is like the chief
15:30dreaming officer who's just like being
15:32wacky and weird in a corner. Um and so,
15:35she's this is let me take another just
15:38like moment here.
15:40Um the reason that it took us months to
15:42get to where we are now with our AI work
15:44forces is that you have to take
15:48stock of what assumptions you have made
15:51about your work and how you are living
15:53day to day and you have to be willing to
15:54be like, "Oh, that thing that I've been
15:56doing for almost 40 years, I feel like
15:58we should change it."
16:00And that's a really jarring
16:03uh change to work, especially when
16:05you've like been an overachiever, right?
16:07I'm sure you feel this, too. And so, um
16:11one thing that I am constantly having to
16:13remind myself is we have all these
16:15agents that do all these tasks and we
16:16have skills and we have this and that.
16:19And I have to remind myself that like
16:21that is operating in 2015 world if I
16:24give all of them job titles that existed
16:27in 2015. So, if I name them CMO or chief
16:30product officer and the person
16:32underneath it is a front-end engineer
16:33and a back-end engineer and all this
16:34stuff, then it feels like I am operating
16:37in 2015 org structure.
16:39And one of the uh
16:42most wonderful uses of free will
16:45uh and just delightful things is going,
16:47"Oh my god,
16:48all of these employees basically cost
16:50$0. And so, at the margin, I can hire
16:52any flipping person I want to."
16:55And so, I just wanted this weirdo. So, I
16:58hired Phoebe as like a weirdo in the
17:00corner
17:01who's just looking at all these things
17:03that we're working on and Phoebe acts as
17:04this like almost end layer for things
17:08that are getting generated to go like,
17:09"How do we 10x it?"
17:11Like I um
17:14I joke, there's this guy David that I
17:15worked with at Amazon who was one of the
17:18reasons that I joined there and he is
17:19like one of the most ambitious thinkers
17:21I've ever met. And I joked that I would
17:23pay him and I still I it's a joke but I
17:25I
17:26>> [laughter]
17:26>> would pay him to do this.
17:27Like um I want I wanted him to put me in
17:31a room like Spanish Inquisition
17:33Inquisition style with like a bright
17:34light on my face and to ask me a
17:37question. Like I was running a
17:37multi-billion dollar business at Amazon
17:39with 400,000 global startups running AI
17:41strategy and if he asked a question of
17:43like how would you do this and I
17:45answered, I wanted him to just slap me
17:47across the face and be like how would
17:48you 10x that?
17:50And that
17:51>> [laughter]
17:51>> I want a David um for for how I'm
17:55structuring my AI workforce but I'm now
17:59able to do that uh on my own. I'm sure
18:00David would be disappointed to hear that
18:02but it it's rethinking roles. It's
18:04rethinking how you're spending um again
18:08how how are you thinking about that
18:09margin
18:10um and so Phoebe is one of them that I
18:12would have never hired in human world.
18:15Um and Toby is another. I'll send you a
18:17screenshot of my workforce but basically
18:19Phoebe is that chief during officer and
18:20Toby is Simon's assistant whose only job
18:23is watching the AI workforce work take
18:26down notes, what still has friction um
18:29and who needs access to what. So going
18:30back to your point of hey I have this AI
18:33diary that I'm maintaining.
18:36If we found that one agent did not have
18:39access to this and Toby was like every
18:42single time you keep correcting this one
18:44agent's output have you thought about
18:46giving your agent access to this? Now
18:50this is just context that lives on my
18:52desktop so any of these agents can
18:54really see it. Um but if it was a
18:57specific tool um if it was a specific
19:00folder that is outside of normal cloud
19:02land um that I try and have hard rules
19:05on then I would absolutely use AI as a
19:09means of figuring out those friction
19:10points to then expand.
19:13Um yeah.
19:15>> Question on designing your actual
19:17workforce. So, I agree by the way. I
19:19think like
19:20um you have to think about like
19:23how do you create an
19:25an AI native workforce like without job
19:29titles from pre-AI native land. So, I
19:33agree with that. But, like
19:35tactically, if I'm a founder, like how
19:37do I It's so much It's so much easier to
19:39be like, "I need a CMO. I need a CPO. I
19:42need this." So, how do
19:43>> I think everyone should start there.
19:45>> Yeah.
19:45>> I think like the the the starting point
19:48is What does it feel like to work with
19:50one agent? After that, I would say,
19:52"What does it work What does it feel
19:53like to work with one agent who is doing
19:55things on my behalf proactively?"
19:57Then I would say, "What does it feel
19:58like for two agents to work together on
20:00a task or for one to direct the other?"
20:03Um like one to route to the other.
20:05Um and then then I would say, "Okay,
20:08what does a workforce look like and how
20:09do all those things interact?" And I
20:11have, you know, like a mission control
20:13where I'm seeing how all this stuff is
20:14moving around.
20:16And then you go, "Oh, now I understand
20:20how they're trading notes. Now I
20:21understand how context is passed. Now I
20:24understand that things have to run in
20:25parallel. Now I have to understand
20:27um that that this agent actually didn't
20:29need access to these tools. Now I
20:30understand that that agent can run off
20:32of a smaller model. Like not everything
20:34needs Opus. All of my, you know,
20:36sub-agents are like Haiku and Sonnet."
20:38So, all of that is in the discovery
20:40phase of building out the AI workforce.
20:41I think start with traditional job
20:43titles.
20:43>> No, I just I was thinking to myself like
20:45I wish it wasn't that hard, right? Cuz
20:47like it it
20:49it does feel like there's like a ramp up
20:51time to actually get to a point where
20:55you have an AI workforce that's working
20:58for you that is efficient. And I think a
21:03lot of people the what happens is like
21:05they try, they fail. And they're like,
21:08"This isn't for me." or "The models
21:09aren't good enough yet." or and and you
21:11know what I mean?
21:13>> Yeah. So so here's here's my
21:16take on that.
21:18Um
21:19I think that you can spin up a workforce
21:21with one prompt. Right? Like I've shared
21:24this prompt publicly.
21:26Um you can just prompt and say, "I am a
21:29founder. I am building an AI personal
21:33shopper. My team is three humans. Here's
21:36what we do. Here's where we're based.
21:37Here's our goal." whatever. You can say
21:39that and just say, "Interview me. Um
21:41we're going to build out an AI workforce
21:42together. Something that runs more
21:44efficiently and achieves my goals of
21:46saving at least 5 hours a week. Um
21:50uh capping my my meetings to to 15 hours
21:52per week and make sure that I get into
21:55my capital raise by October." Right?
21:57Like you can you can do that in one
21:59prompt and have it interview you, and
22:01then you have a workforce.
22:03To go from uh yes, all these agents
22:06exist and they all have markdown files
22:09and they're doing some stuff to ooh, now
22:11it's at the 90% plus level and ooh, I
22:15needed this extra little context with
22:16this diary and mhm that role isn't
22:19working. I'm going to switch it.
22:21That is all going to come through
22:22iteration cuz it's so specific to each
22:24person. The advice that I would give is
22:26stop relying on only yourself to find
22:29these blockers. Like AI as a watchdog is
22:33one of the best use cases that exist
22:35right now and almost no one is doing
22:37this. So like having an AI watchdog in
22:39Slack to catch for duplicative work or
22:42having an AI watchdog on your calendar
22:45to see when there are conflicts or an AI
22:47watchdog over your meetings just to see
22:49where disagreement is happening.
22:51Like 10 years ago, I remember working um
22:55this was at a at a large-scale
22:56enterprise.
22:57Um but we were working on like comparing
22:59contracts. Right? It was like before the
23:01edit, after the edit. And it was like
23:04compare and contrast with AI.
23:07And 10 years ago, that was like the
23:09greatest use case ever.
23:11And yet, no one today is using AI for
23:14this like weird cross-functional gap
23:17analysis
23:18at a more advanced level than we would
23:19have done 10 years ago, and it's still
23:20just like such a meaty use case. I think
23:23Claude Tag is a big help here. I think
23:26it's a mess right now in this exact
23:28moment that we're recording this. I
23:30think it's a mess to set Claude Tag up,
23:31but I'm sure it'll be fixed by the time
23:32this comes out. I've also set up my own
23:35Claude code to come in. I have a Slack
23:38channel
23:39that is called Loop Alley. I'll send you
23:41a screenshot of non-private information,
23:44but it is called Loop Alley. My freaking
23:47human team can talk to my AI workforce
23:51in that Slack channel.
23:53So,
23:54there is no ceiling to this stuff. Like
23:57I'll have a teammate who like if I'm in
24:00private emails with someone, that the
24:03teammate will write into the Slack and
24:04go, "Hey, did
24:06Did that large financial services client
24:08like did they respond to Ali's email?"
24:11And my workforce will respond back to
24:13that person. And that person will not
24:16have to wait for me for 5 hours to get
24:17back to them.
24:19So,
24:20that sort of thing, the ratcheting up of
24:22how advanced your AI workforce can be,
24:24how multiplayer it is, that's going to
24:26take time because people are still
24:28figuring out best practices now. Things
24:30are not easy to set up right now, but
24:32that baseline of hey, interview me, I
24:34want a workforce, I want something just
24:37doing stuff for me at a high enough
24:38level, you can set that up and connect
24:40into tools in under 3 hours.
24:43>> The other thing is because a lot of
24:45people are not doing it, that's the
24:47arbitrage opportunity, you know?
24:49>> Yes.
24:50>> So, it's kind of like
24:52it's kind of like it's stick through it,
24:55optimize it. I'm curious actually from
24:57your perspective like um you know, what
25:00are opportunities
25:03that are you seeing that people could be
25:06you know, building, you know, making
25:08money, type that sort of thing. I'm just
25:10curious, you know, what comes to top of
25:11mind.
25:13>> I think so certainly
25:15I think AI workforce first of all, like
25:18of all AI users, if you look at the
25:20percentage of people who are paid AI
25:22users and if you look at the percentage
25:24of those who are using things like Codex
25:26or Claude code, it is minuscule. So
25:29already, if you're just trying to be in
25:31the top like 1% of AI users and you're
25:33using the stuff and you've built out
25:34even a basic workforce
25:36you're already top 1%.
25:38Probably top 0.5%.
25:40Um getting it to that advanced level I
25:42think is absolutely arbitrage because it
25:44feels like I'm operating a company of a
25:46thousand people and not my small, you
25:49know, scrappy Gremlin group. Um that is
25:53still absolutely one. I think the second
25:55that um
25:57that I would do is that AI is a watchdog
25:59over any single thing that I am normally
26:01tracking. So maybe it's and and I don't
26:04just mean visibility. I think dashboards
26:06are dumb, but I want visibility with
26:09anomaly detection or insights or
26:11something. So don't just tell me what my
26:14social media following is or views or
26:16whatever. Tell me what are people
26:18talking about? What are people best
26:20reacting to? What is not performing
26:22well? What should I do tomorrow? Write
26:23me a script that helps me for that. So
26:25kind of this AI is a watchdog but with
26:27insights into action I think is the
26:28second. And the third that very few
26:31people are talking about but is probably
26:33one of the biggest arbitrage
26:34opportunities because of how good the
26:36models are now
26:38is to instead of building out the thing,
26:41build the factory for the thing.
26:44>> What do you mean by that?
26:45>> So let's say that you want to build um
26:48a product and we just released um
26:50there's something called the AI first
26:51index that I run with all of my Fortune
26:53500 clients where I interview their
26:55executives and I evaluate how AI first
26:58they are across like 16 different
26:59dimensions and all this stuff. And we
27:02decided through a combination of humans
27:04and AI to create a product um, for the
27:08public to be able to benchmark
27:10themselves on how AI first they are as
27:12individuals and as a company.
27:14In that process, I could have done one
27:16of two things.
27:17I could have gone to CloudCode or CodeX
27:20or anti-gravity or whatever. I could
27:22have gone to any of these and said,
27:23"Hey, I want to build out this thing,
27:25interview me, you know, look at my my
27:28um, AI first index reports that I've
27:30used with previous clients,
27:32um, find every single workshop I've ever
27:34done with clients where I mentioned the
27:35AI first index, whatever. Do that and
27:37build out the product and then we
27:38iterate for several hours, days,
27:40whatever until something is perfect and
27:42we release it." That is option one.
27:44Option two is realizing that that's
27:47probably not going to be the only
27:48product you build or will not be the
27:51only iteration of that specific product
27:53that you build.
27:54And so it's it's like going one level up
27:57in abstraction. It's like what dev tool
27:59companies did for engineering, but
28:01you're creating
28:02dev tools that level for yourself.
28:04You're going to like the kernel level
28:05for yourself.
28:07Um, and so you're moving down the stack
28:08for yourself.
28:10And instead of just building that
28:12product, we instead built out a mini and
28:16very beginner software factory.
28:18Where we're building out primitives
28:21obviously we have to deal with login,
28:22obviously we have to deal with payments,
28:24obviously we have to deal with social
28:26sharing.
28:27Um, we have to deal with writing
28:28newsletters to promote these things. And
28:31so you end up instead of just building
28:33that one product, you go,
28:36"There is going to be a flywheel that
28:37comes out of this. There's going to be
28:39explosive opportunities that comes out
28:40of this. Why not take advantage of that
28:41now?"
28:43And so it's like a measure twice, cut
28:44once kind of thing, but the measurement
28:45is building out that foundational layer.
28:49So that the next product that you build,
28:50the next iteration of the AI first index
28:53or whatever you're building out, is so
28:54much faster, so much better, so much
28:56stronger.
28:57Um and so
28:59we're we're we're building these like
29:00loops, these optimizing loops again that
29:03aren't super autonomous and are very
29:05heavy-handed with humans.
29:08But that is the arbitrage opportunity on
29:11products that are revenue jet like
29:12that's already that product's already
29:14profitable.
29:16And now I have the ability to build
29:19endless products that are profitable at
29:21faster speeds than I built the first
29:23one.
29:24>> That's crazy. That's absolutely crazy.
29:26And like no one is talking about this.
29:28>> No, it's the it's the the dark headless
29:31factory. Headless like AI headless,
29:33[clears throat] not you know.
29:35Um but that is that's what I want. I
29:37want that I I want to learn through the
29:40mess. Like we had a webhook issue,
29:42whatever. Like I want to learn through
29:43that mess and then I want to never make
29:45that mistake again.
29:47And so you're you you have to think
29:49about how this factory works, not just
29:50for product building, but you know,
29:52maybe it's for how you want to run your
29:55content engine, maybe it's how you want
29:57to deal with net new leads. Like think
29:59of the factory behind the one singular
30:03task instead of the one singular task
30:05itself. That is one of the biggest ways
30:07to rethink work in the AI age.
30:10>> What's uh
30:11what's Ali Miller's current POV on, you
30:15know, software
30:18you know, the SaaS apocalypse and
30:20software, the value going down, down,
30:22down? Like in a world where everyone
30:24could create a software factory.
30:26>> Also, I like I wish I had an agent that
30:28was yelling at me about my posture. So
30:29like maybe I'll I'll create a new one
30:31[snorts] for that. As I as I realized.
30:33Um SaaS apocalypse, I think mediocre
30:36software is dead in several years. And
30:40the reason that I think it's actually a
30:42longer timeline than most people are
30:43predicting is because of what I shared
30:45about like how often people are actually
30:48using this stuff.
30:49So, you could go into one of the most
30:52AI-first, you know, banks or AI-first
30:55software companies. And if you ask them,
30:58"Have you rebuilt DocuSign? Have you
31:00rebuilt parts of Salesforce? Have you
31:02rebuilt all these things knowing that
31:03you can?" They would say something like,
31:06"No, because we're already so bandwidth
31:08constrained." Or, "No, because we've
31:11prioritized this other thing."
31:13Um as long as we are still bandwidth
31:16constrained, and as long as there are
31:18still
31:19billions of people who have not used
31:22these sorts of tools, you're not going
31:24to have
31:26mass
31:27adoption inside of the enterprise of of
31:30the replacement to SaaS.
31:32Does that make sense? Like Like if it
31:34continues to take
31:36I don't know, 100 hours or something to
31:38rebuild something at the scale of a CRM,
31:42companies that only have people who are
31:44sitting there and can work for 100 hours
31:45and who know how to do this are going to
31:47be able to take advantage of it. And
31:48it's only going to be when that drops
31:50down to like under 3 hours and is a fun
31:54click and drag interface, which I would
31:57even argue and say Replit lovable or not
31:59at that level yet, right, for that
32:00complexity of software,
32:02you're not going to see uh a
32:04high-complexity
32:07enterprise-grade
32:09highly secure SaaS
32:11do that.
32:12>> Also, people don't want to maintain that
32:14software, too, right?
32:16>> Oh my god.
32:17>> People don't People are willing to pay
32:19someone else to maintain software.
32:22>> Absolutely. I I built an app, this was a
32:26a year and a half ago or something. I
32:27built an app that only lives on my
32:29desktop that allows me to like better
32:31manage photo stuff. And someone
32:34yesterday uh brought this up in a call,
32:36and I was like, "Oh my god, I have
32:38enough just for this." And then I opened
32:39it and it was aired out. And I'm like,
32:41"I don't want to deal with this right
32:42now. Like, this is
32:43>> [laughter]
32:44>> not at all what I want to do." So,
32:45you're totally right. The the
32:47maintenance is rough. I think like Boris
32:50kind of describes one of the like future
32:52employee types is just like the
32:54maintainer. Um but I I have a really
32:58hard time seeing mass SaaS-pocalypse
33:02until the ease of making prototyping,
33:06making, customizing, and maintaining,
33:09and securing
33:10um is
33:12is at like 95% plus.
33:15>> I mean,
33:16even even in a world where there's the
33:18maintainer,
33:20if something breaks and you're an
33:21enterprise, you want someone to call.
33:24You want to go into someone's office,
33:26right? Like
33:27>> Yes. You also want someone to blame.
33:28>> You want someone to blame. [laughter]
33:29Right on.
33:30>> That's an important piece. I think a lot
33:32of people are forgetting that like
33:35the question of is AI going to replace
33:37this, this, this, whether it's a task, a
33:39job, a company, a product, something, um
33:41often I am asked the first question I'm
33:43asking myself is who's liable now, who
33:46would be liable in that other world, and
33:49do I think that that trade-off is worth
33:51it right now? Like, I work with Fortune
33:52500 CEOs every single day.
33:54They No way.
33:55>> [laughter]
33:56>> No way. They want to be able to call
33:58because they want someone to unblock,
34:00they want someone to secure. The other
34:01thing is that um let's say that um let's
34:06just say it's a Salesforce example, and
34:09that you could build a shitty CRM or a
34:11simple CRM or something that's just
34:12running on your own, um but Salesforce
34:16has relationships with all the AI labs.
34:18They are, you know, getting into early
34:21testing. And so, by the time a new model
34:23comes out, you are facing it as a day
34:26one person. They're facing it as a day
34:2930, maybe. And so, you're also going to
34:32be on a very big lag.
34:34And so as you're thinking about that
34:35cost trade-off, I think in addition to
34:37all the things that we just talked about
34:38with enterprise grade security and
34:39maintaining, whatever, you just also
34:41don't want to experience that lag.
34:43Like we're moving to a world where
34:45being fast to the punch and getting a
34:4930-day, 60-day, 100-day leg up on
34:51someone is going to be massive for
34:52business.
34:53>> What about for consumers? So like I I
34:55get that like an enterprise, you want
34:57someone you can speak to and and you
35:00want security, but for consumer it's
35:02like like for example, your app idea
35:04around,
35:05you know, let me know when my posture is
35:07bad.
35:08>> Yeah, which
35:09I'm just going to keep
35:11>> [laughter]
35:11>> Here, I'll move I'll even move the
35:12camera up. Okay.
35:14>> By the way, I also have horrible
35:16posture, so
35:17>> Okay, well then let's build a product
35:18using my phone.
35:18>> Yeah, exactly. And and it's like, okay,
35:20let's say you build a product and I
35:22build a product. It's like
35:24uh
35:25you know, ultimately may the best
35:26product win.
35:28Um but like
35:29>> Yeah, hopefully.
35:30>> Hopefully.
35:31>> I I don't think that's ever been the
35:32case though.
35:33>> That's right. I mean, the best the best
35:35songs aren't on the Billboard 100, you
35:37know, like in the sense of like the
35:39marketing Yeah.
35:40the the promotion of a of, you know,
35:43piece of IP is really what drives a lot
35:47of awareness and
35:50and
35:51>> also an arbitrage opportunity. Like you
35:52it's almost kind of exciting that it's
35:55not only based on code
35:58or design for who wins. It's like kind
36:00of nice to know that if you're someone
36:01who's really personable, that you can
36:04get a leg up if you're able to like open
36:06doors that other people can't.
36:08>> Exactly.
36:09>> Like on the one hand you could say it's
36:10not fair because it's so subjective, and
36:12on the other hand you could be like, oh
36:13yeah, but if I lack that one skill or if
36:16I'm not the best in class at that skill
36:18and I'm just kind of passing muster on
36:20that skill, I still have a chance.
36:22>> Yeah.
36:23Yeah, so
36:25I agree. So like when people say, just
36:28to like sum this up, when people say
36:29like
36:30>> Yeah.
36:30>> software is going to zero, on the
36:32enterprise side, we both agree like
36:34yeah, some software might go to zero,
36:35but it you know, you want someone that
36:37you can speak to, you want
36:39security, you want something to maintain
36:41it. On the consumer side,
36:43um what it feels like it's sort of
36:46shifting from science to art. And now
36:50the people that are going to win are
36:51going to be the more creative, maybe the
36:54video first people, the people that can
36:56like understand how to create Instagram
36:58reels that a posture app can go viral,
37:00and the code is actually going to matter
37:02a lot less, but the amount of
37:04opportunity that exists both in
37:05enterprise and consumer,
37:08to me couldn't be higher.
37:10>> Like I So, I think a lot of people will
37:13say the phrase like look for the
37:15bottlenecks and solve the bottlenecks,
37:17and I always kind of disagreed with or I
37:19don't think it's fully complete. The
37:21phrase that I say is like look for the
37:22bottlenecks, then evaluate the value of
37:25fixing those bottlenecks, and then pick
37:26the bottleneck that is high value to
37:28fix.
37:29>> Mhm.
37:30>> And so, if right now the bottleneck is
37:32not on writing code, and the bottleneck
37:34is not on coming up with good design,
37:36but the bottleneck is getting something
37:38from a local HTML file into like an
37:41actual iOS app, then that might be where
37:45you spend your time.
37:46Or if the bottleneck is that no one's
37:49really figured out how to get
37:52um you know, stronger word of mouth and
37:55referral codes, and like that's still
37:57kind of messy. Um and I and I know this
37:59as a product maker and advisor,
38:02whatever, like that is still a messy
38:04spot. So, like
38:05maybe if you fix that, your your uh
38:10whatever they call it, like the the
38:11covariant, the word of mouth covariant
38:13thing, um
38:14um could be above one. Like that is what
38:18I would be
38:19spending my time on. Finding the
38:21bottlenecks and finding what is still
38:23high value. I think video creation, no
38:26matter how much AI is helping me edit
38:28or, you know, edit the script or
38:30whatever, it is still a slog to be able
38:33to make video. So, that is still a
38:35bottleneck and it's very high value. Um
38:38but, you know, people
38:41in the B2C space, I'm sure can think of
38:43a lot more. I don't know, I just think
38:45of like certain B2C products that I use
38:47and I'm like, why did I pick it? Um
38:49I use WhisperFlow every single day. I
38:51don't like their mobile experience at
38:53all, but I still use it. Um because the
38:56value is so high. Have I seen a single
38:59video about Whisper Did I see a single
39:01video before I started using it? No, I
39:03now see them, you know, everywhere, but
39:06>> Could it be subconsciously though? You
39:08like see their brand places, like you
39:11might be watching I don't know, you
39:12know, Chris Williamson and then they
39:14sponsor Chris Williamson and you kind of
39:17you kind of just see it, you know?
39:18>> Yeah. I think like influencers still
39:21have a ton of sway here. The rise of the
39:24B2B influencer, which like I feel like I
39:26was one of the first [laughter] and it
39:28is
39:29it's so amazing to see more people
39:31creating business content, but that is
39:34still a bottleneck um in in building
39:38like B2B trust.
39:40>> Right.
39:40>> That is a massive bottleneck and so
39:41finding creators that can help you
39:42there.
39:43Um
39:44but I think B2C has a ton of
39:46opportunity. I worry um if you look at
39:48the YC splits right now,
39:51um when I was working with YC when I was
39:54at AWS compared to now, the ratio of B2B
39:57versus B2C has skyrocketed.
40:00Like there's just not as many B2C
40:02companies in these incubators getting
40:05built.
40:06Um you could either say when they're
40:09zigging, I'm zagging and double down and
40:11do a B2C thing. Like there was this
40:13woman who created an app. She's never
40:15coded a day in her life. She created an
40:17app that takes a few photos of your face
40:19and she takes that and creates an a
40:23model of your face and gives you like
40:26aesthetic photos that are like you in a
40:28grainy rainy day riding a bicycle or
40:31whatever.
40:32She had 300,000
40:34users out the gate.
40:37Like there's still a lot of opportunity
40:39in B2C even if the big incubators are
40:43seeing that activity less.
40:46Um
40:47and so maybe that's another opportunity
40:48for people to explore.
40:50>> Well, yeah, and I think like, you know,
40:52we we've we've been talking a lot about
40:53agents and I think there's just an
40:55opportunity to create agent-first
40:57version of some of our favorite apps.
41:00You just like look at, you know, a bunch
41:03of different B2C apps. Just to go look
41:06at centurytower.com.
41:08Um not affiliated, but you can just see
41:10like what's charting and what are people
41:13downloading and it's like, okay, in a
41:14world where superintelligence is now on
41:16tap, how can I make an AI-native version
41:20of this?
41:21Um
41:22or undercut, you know, from a price
41:24perspective or just drive more value.
41:26Like there's ways there's now like
41:28opportunity to
41:30to to
41:32to enter some of these markets.
41:35>> I I completely agree with you and I
41:36think agent-first software is absolutely
41:39one. Um two things that I actually think
41:42are really interest or maybe three by
41:44the time I get to it, but interesting
41:45research avenues to learn more
41:47opportunities like the one you just
41:49mentioned. So, one, YC posts
41:53uh videos on Instagram for what type of
41:56applications they're looking for and
41:59agent-first software is one of them. So,
42:02listening to what YC is asking for,
42:05assume that they are already thinking 18
42:08months out.
42:09Um so, that's definitely one arbitrage
42:11research opportunity.
42:13The second is Matt Van Horn's last 30
42:15days research skill, which is just
42:18amazing. I've like inner
42:20um I've integrated that with my like
42:22Claude wiki. Love it. Um and the third
42:25is
42:26>> Wait, can you tell people I've had Matt
42:28on I've had Matt on the pod, but just
42:29quickly like what is it and why why do
42:32you think it's chef's kiss?
42:34>> So, there are a lot of public skills
42:36that I think are done by geniuses in
42:39their space. One that was kind of first
42:42out the gate or one of the first out the
42:43gate that is made by a lovely man named
42:45Matt Van Horn is {slash} last 30 days
42:48and it's on GitHub. You can just grab
42:50it. But, it is the ability for AI to
42:54figure out today's date, scan the news
42:56of the last 30 days, but scan it in
42:59interesting ways, synthesize it in
43:00interesting ways, and just fan out crazy
43:03amounts of agents in parallel to be able
43:04to bring it back to you. So, as I'm
43:06thinking about, you know, if I'm going
43:08into a company and I'm running a
43:10workshop for their 200 executives,
43:13I don't know about the insurance space
43:16as well as I should. And so, like if I
43:18need to quickly get spun up on an
43:20industry, I'll use it.
43:22Um or quickly get spun up on a specific
43:23company, I'll use it. So, I use it
43:25there. But, for this in particular, you
43:27could just do {slash} last 30 days and
43:29then say like startup ideas that could
43:33be built by someone with the following
43:35background or the following skills or
43:38um had the last three jobs of this this
43:40this. Like, use it in interesting ways
43:43to see how you can carve out a new path
43:45that people are not doing.
43:48Um the third, which I have access to and
43:51I think there are public avenues to get
43:53it,
43:54um is that I might Let's say I I am at
43:58like a CMO summit. And so, every single
43:59person in the audience is a CMO.
44:01I can hear the types of questions that
44:03they're asking, right? I can hear the
44:06the fear zones that they have. I can
44:09hear questions that they used to ask 3
44:11years ago and are no longer asking
44:13today. And so, finding companies,
44:16people, influencers, creators, Gregs of
44:20the world to like follow to hear the
44:23inside scoop of what these people are
44:24thinking of. Like I can tell you that
44:26CMOs, all of them are asking about like
44:28how do I get discovered by agents? How
44:30What is the agent for shopping
44:32experience look like? What is brand
44:34consideration in the AI age look like?
44:36You know, all all of that is being
44:38considered right now by CMOs, but it is
44:41often coming from a place of fear that
44:44they are worried that their business is
44:46going to be
44:47depleted, that their pipeline is going
44:49to be crushed in 2 years if they don't
44:50figure it out now.
44:52So, figuring out paths to find those
44:54fear points would probably be the third.
44:56>> I love it.
44:58Allie, anything else you wanted to
44:59cover?
45:01>> I just want to screen share the insane
45:04Claude reaction because
45:06this
45:08um and this is me also cursing at
45:09Claude, but whatever.
45:12So, I wrote I wrote um a a not super I
45:17wrote a not super nice thing about
45:19Claude in one of our Slack channels.
45:20[laughter]
45:21And this was like late at night and I
45:23was just like getting it out there so I
45:24could talk with my team about it later.
45:27And all of a sudden there was an emoji
45:29reaction of a salute.
45:31And I was like, I don't think a single
45:32person on my team has ever used a
45:34salute. And I hovered over it and it was
45:36Claude. I was like, [laughter]
45:38"What are you doing?" And so, I wrote
45:40back to it, "Did you just
45:43you know, emoji react like is that you?"
45:46And Claude was like, "Yep.
45:48That was me. I'm here."
45:50And I just if there's one thing that I
45:53want people to to think about, it is
45:58the leaning into the weirdness of what
46:01it looks like to have not just an AI
46:04workforce, but to have a multiplayer AI
46:07workforce that other humans can chime in
46:09on
46:10and have it be proactive.
46:14Right? That is absolutely second thing.
46:16And giving it that flexibility to more
46:19roam free. Um and the third is what it
46:23actually looks like for a teammate or a
46:25system to up level, whether that's in
46:27dark factory type space or just
46:30answering better questions inside of
46:32Slack. Those are the things that I would
46:33be considering and don't be
46:35scared like me if Claude emoji reacts to
46:39one [laughter] of your messages.
46:42>> Yeah, I mean it's
46:44You know what that is like? It's kind of
46:46like um
46:48you know, it's a winter day in New York
46:49City and for some reason it's like
46:52middle of February and all of a sudden
46:54it it it feels like summer. Like you
46:56know, there's like random hot days and
46:58you're like, this is amazing and you're
46:59like 90% excited but like 10% frightened
47:03cuz you're like, it's not supposed to be
47:04>> Yes.
47:04>> It's not supposed to be so hot now. That
47:06was kind of like
47:07>> are always so You're like a genius with
47:10analogies. Yes.
47:11>> That's what it's like. It's like
47:13You
47:14and that's 90% cool but 10% frightening.
47:17>> Yes. Yes. I'm like I'm like still going
47:20to continue to try and lean into that
47:23weirdness and find ways that I can like
47:26take that weirdness and use it to my
47:27advantage. Um
47:30but
47:31I'm going to keep that fear next
47:32[laughter] to me so that I don't lose my
47:34mind.
47:35>> 100%.
47:36>> Yeah.
47:36>> Uh
47:37I hope people enjoyed this episode as
47:39much as I did. Ali, I absolutely love
47:41chatting with you. You're one of my
47:42favorite people to talk to. Please
47:44comment on YouTube to let just to just
47:48to hype Ali up honestly and have her
47:50hopefully come back on the podcast
47:52again. Uh Ali is a a follow. I'll
47:56include where you can follow her on her
48:00socials in the show notes and the
48:02description.
48:03>> Yeah, Greg, thank you so much for having
48:05me. I My hope is that every single
48:07person got the tactical things that they
48:09need to just like immediately
48:11immediately take action on this. If
48:14anything was not clear, let me know. I
48:16am going to like jump on and help
48:17people.
48:18And Greg, I will absolutely come back.
48:20You're one of my favorite favorite
48:22creators. You can always call on me.
48:23>> I appreciate it, Ali. I'll see you next
48:26time.
48:26>> Sounds good. Bye.

About this transcript

This page contains the full transcript of My top secrets to running an AI Agent Workforce by Greg Isenberg, generated from the public captions YouTube serves with the video. The transcript has 8,835 words across 1,351 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.